AI Self-Improvement May Take Longer Than Expected: What Does New Research Show?

According to a new academic study, recursive self-improvement—one of the artificial intelligence industry's grandest promises, enabling systems to completely enhance themselves without human intervention—may be much further off in the future than previously anticipated. Led by Princeton University, the recent research reveals that while current AI agents can successfully execute engineering tasks, they still lack the creativity, reasoning, and judgment required to conduct open-ended scientific research.

How Was the New Research Conducted, and What Is "Shadow Evaluation"?

A multi-institutional research group spearheaded by Peter Kirgis and Sayash Kapoor of Princeton University developed a novel method called "shadow evaluation" to test AI models' capacity for high-level research. As part of this test, Anthropic’s Claude Opus 4.8 model was tasked with solving the research questions of two high-quality papers submitted to NeurIPS 2026, a prestigious machine learning conference, that had not yet been made public.

The agents were provided with six days, a $3,000 API credit, a GPU budget, virtual machines, and internet access. This setup prevented the models from simply memorizing from their training data, strictly testing their capacity for genuinely original scientific reasoning.

Engineering Success, Lacking Scientific Creativity

Following the test, the human authors of the original papers rejected both studies generated by the artificial intelligence. The research findings clearly highlight both the AI's strong performance in the technical and practical aspects of the process and its shortcomings in strategic decision-making:

• Areas of Success: The agents excelled at engineering-heavy tasks such as conducting literature reviews, automatically running hundreds of experiments, writing code, and compiling results.
• Areas of Shortcoming: The models struggled with selecting the right hypotheses, anticipating which evidence would resolve a problem, avoiding wasted time on meaningless or minor synthetic datasets, and delivering the innovative, publishable contributions unique to the field.

Industry Implications and AI Development Timelines

The speed of AI models in narrow technical tasks—such as coding, synthetic data generation, and chip optimization—had fostered a perception within the industry that a fully autonomous self-improvement loop was imminent. However, this new study demonstrates that the judgment, taste, and open-ended problem-solving demanded by high-level scientific research processes cannot be immediately replaced by current architectures. This suggests that speculative timelines promising exponential progress in AI must be re-evaluated against concrete evidence.

*This news report has been prepared based on data published by MIT Tech Review — AI.